change table name
This commit is contained in:
parent
c0cd08d45e
commit
f395dd9ca0
2 changed files with 73 additions and 73 deletions
128
ggml-common.h
128
ggml-common.h
|
@ -1022,70 +1022,70 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
|
|||
0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
|
||||
GGML_TABLE_END()
|
||||
|
||||
GGML_TABLE_BEGIN(uint32_t, i2_q8, 256)
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00010000, 0x01010000, 0x00010000, 0xff010000,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00ff0000, 0x01ff0000, 0x00ff0000, 0xffff0000,
|
||||
0x00000100, 0x01000100, 0x00000100, 0xff000100,
|
||||
0x00010100, 0x01010100, 0x00010100, 0xff010100,
|
||||
0x00000100, 0x01000100, 0x00000100, 0xff000100,
|
||||
0x00ff0100, 0x01ff0100, 0x00ff0100, 0xffff0100,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00010000, 0x01010000, 0x00010000, 0xff010000,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00ff0000, 0x01ff0000, 0x00ff0000, 0xffff0000,
|
||||
0x0000ff00, 0x0100ff00, 0x0000ff00, 0xff00ff00,
|
||||
0x0001ff00, 0x0101ff00, 0x0001ff00, 0xff01ff00,
|
||||
0x0000ff00, 0x0100ff00, 0x0000ff00, 0xff00ff00,
|
||||
0x00ffff00, 0x01ffff00, 0x00ffff00, 0xffffff00,
|
||||
0x00000001, 0x01000001, 0x00000001, 0xff000001,
|
||||
0x00010001, 0x01010001, 0x00010001, 0xff010001,
|
||||
0x00000001, 0x01000001, 0x00000001, 0xff000001,
|
||||
0x00ff0001, 0x01ff0001, 0x00ff0001, 0xffff0001,
|
||||
0x00000101, 0x01000101, 0x00000101, 0xff000101,
|
||||
0x00010101, 0x01010101, 0x00010101, 0xff010101,
|
||||
0x00000101, 0x01000101, 0x00000101, 0xff000101,
|
||||
0x00ff0101, 0x01ff0101, 0x00ff0101, 0xffff0101,
|
||||
0x00000001, 0x01000001, 0x00000001, 0xff000001,
|
||||
0x00010001, 0x01010001, 0x00010001, 0xff010001,
|
||||
0x00000001, 0x01000001, 0x00000001, 0xff000001,
|
||||
0x00ff0001, 0x01ff0001, 0x00ff0001, 0xffff0001,
|
||||
0x0000ff01, 0x0100ff01, 0x0000ff01, 0xff00ff01,
|
||||
0x0001ff01, 0x0101ff01, 0x0001ff01, 0xff01ff01,
|
||||
0x0000ff01, 0x0100ff01, 0x0000ff01, 0xff00ff01,
|
||||
0x00ffff01, 0x01ffff01, 0x00ffff01, 0xffffff01,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00010000, 0x01010000, 0x00010000, 0xff010000,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00ff0000, 0x01ff0000, 0x00ff0000, 0xffff0000,
|
||||
0x00000100, 0x01000100, 0x00000100, 0xff000100,
|
||||
0x00010100, 0x01010100, 0x00010100, 0xff010100,
|
||||
0x00000100, 0x01000100, 0x00000100, 0xff000100,
|
||||
0x00ff0100, 0x01ff0100, 0x00ff0100, 0xffff0100,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00010000, 0x01010000, 0x00010000, 0xff010000,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00ff0000, 0x01ff0000, 0x00ff0000, 0xffff0000,
|
||||
0x0000ff00, 0x0100ff00, 0x0000ff00, 0xff00ff00,
|
||||
0x0001ff00, 0x0101ff00, 0x0001ff00, 0xff01ff00,
|
||||
0x0000ff00, 0x0100ff00, 0x0000ff00, 0xff00ff00,
|
||||
0x00ffff00, 0x01ffff00, 0x00ffff00, 0xffffff00,
|
||||
0x000000ff, 0x010000ff, 0x000000ff, 0xff0000ff,
|
||||
0x000100ff, 0x010100ff, 0x000100ff, 0xff0100ff,
|
||||
0x000000ff, 0x010000ff, 0x000000ff, 0xff0000ff,
|
||||
0x00ff00ff, 0x01ff00ff, 0x00ff00ff, 0xffff00ff,
|
||||
0x000001ff, 0x010001ff, 0x000001ff, 0xff0001ff,
|
||||
0x000101ff, 0x010101ff, 0x000101ff, 0xff0101ff,
|
||||
0x000001ff, 0x010001ff, 0x000001ff, 0xff0001ff,
|
||||
0x00ff01ff, 0x01ff01ff, 0x00ff01ff, 0xffff01ff,
|
||||
0x000000ff, 0x010000ff, 0x000000ff, 0xff0000ff,
|
||||
0x000100ff, 0x010100ff, 0x000100ff, 0xff0100ff,
|
||||
0x000000ff, 0x010000ff, 0x000000ff, 0xff0000ff,
|
||||
0x00ff00ff, 0x01ff00ff, 0x00ff00ff, 0xffff00ff,
|
||||
0x0000ffff, 0x0100ffff, 0x0000ffff, 0xff00ffff,
|
||||
0x0001ffff, 0x0101ffff, 0x0001ffff, 0xff01ffff,
|
||||
0x0000ffff, 0x0100ffff, 0x0000ffff, 0xff00ffff,
|
||||
GGML_TABLE_BEGIN(uint32_t, i2s_i8s, 256)
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00010000, 0x01010000, 0x00010000, 0xff010000,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00ff0000, 0x01ff0000, 0x00ff0000, 0xffff0000,
|
||||
0x00000100, 0x01000100, 0x00000100, 0xff000100,
|
||||
0x00010100, 0x01010100, 0x00010100, 0xff010100,
|
||||
0x00000100, 0x01000100, 0x00000100, 0xff000100,
|
||||
0x00ff0100, 0x01ff0100, 0x00ff0100, 0xffff0100,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00010000, 0x01010000, 0x00010000, 0xff010000,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00ff0000, 0x01ff0000, 0x00ff0000, 0xffff0000,
|
||||
0x0000ff00, 0x0100ff00, 0x0000ff00, 0xff00ff00,
|
||||
0x0001ff00, 0x0101ff00, 0x0001ff00, 0xff01ff00,
|
||||
0x0000ff00, 0x0100ff00, 0x0000ff00, 0xff00ff00,
|
||||
0x00ffff00, 0x01ffff00, 0x00ffff00, 0xffffff00,
|
||||
0x00000001, 0x01000001, 0x00000001, 0xff000001,
|
||||
0x00010001, 0x01010001, 0x00010001, 0xff010001,
|
||||
0x00000001, 0x01000001, 0x00000001, 0xff000001,
|
||||
0x00ff0001, 0x01ff0001, 0x00ff0001, 0xffff0001,
|
||||
0x00000101, 0x01000101, 0x00000101, 0xff000101,
|
||||
0x00010101, 0x01010101, 0x00010101, 0xff010101,
|
||||
0x00000101, 0x01000101, 0x00000101, 0xff000101,
|
||||
0x00ff0101, 0x01ff0101, 0x00ff0101, 0xffff0101,
|
||||
0x00000001, 0x01000001, 0x00000001, 0xff000001,
|
||||
0x00010001, 0x01010001, 0x00010001, 0xff010001,
|
||||
0x00000001, 0x01000001, 0x00000001, 0xff000001,
|
||||
0x00ff0001, 0x01ff0001, 0x00ff0001, 0xffff0001,
|
||||
0x0000ff01, 0x0100ff01, 0x0000ff01, 0xff00ff01,
|
||||
0x0001ff01, 0x0101ff01, 0x0001ff01, 0xff01ff01,
|
||||
0x0000ff01, 0x0100ff01, 0x0000ff01, 0xff00ff01,
|
||||
0x00ffff01, 0x01ffff01, 0x00ffff01, 0xffffff01,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00010000, 0x01010000, 0x00010000, 0xff010000,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00ff0000, 0x01ff0000, 0x00ff0000, 0xffff0000,
|
||||
0x00000100, 0x01000100, 0x00000100, 0xff000100,
|
||||
0x00010100, 0x01010100, 0x00010100, 0xff010100,
|
||||
0x00000100, 0x01000100, 0x00000100, 0xff000100,
|
||||
0x00ff0100, 0x01ff0100, 0x00ff0100, 0xffff0100,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00010000, 0x01010000, 0x00010000, 0xff010000,
|
||||
0x00000000, 0x01000000, 0x00000000, 0xff000000,
|
||||
0x00ff0000, 0x01ff0000, 0x00ff0000, 0xffff0000,
|
||||
0x0000ff00, 0x0100ff00, 0x0000ff00, 0xff00ff00,
|
||||
0x0001ff00, 0x0101ff00, 0x0001ff00, 0xff01ff00,
|
||||
0x0000ff00, 0x0100ff00, 0x0000ff00, 0xff00ff00,
|
||||
0x00ffff00, 0x01ffff00, 0x00ffff00, 0xffffff00,
|
||||
0x000000ff, 0x010000ff, 0x000000ff, 0xff0000ff,
|
||||
0x000100ff, 0x010100ff, 0x000100ff, 0xff0100ff,
|
||||
0x000000ff, 0x010000ff, 0x000000ff, 0xff0000ff,
|
||||
0x00ff00ff, 0x01ff00ff, 0x00ff00ff, 0xffff00ff,
|
||||
0x000001ff, 0x010001ff, 0x000001ff, 0xff0001ff,
|
||||
0x000101ff, 0x010101ff, 0x000101ff, 0xff0101ff,
|
||||
0x000001ff, 0x010001ff, 0x000001ff, 0xff0001ff,
|
||||
0x00ff01ff, 0x01ff01ff, 0x00ff01ff, 0xffff01ff,
|
||||
0x000000ff, 0x010000ff, 0x000000ff, 0xff0000ff,
|
||||
0x000100ff, 0x010100ff, 0x000100ff, 0xff0100ff,
|
||||
0x000000ff, 0x010000ff, 0x000000ff, 0xff0000ff,
|
||||
0x00ff00ff, 0x01ff00ff, 0x00ff00ff, 0xffff00ff,
|
||||
0x0000ffff, 0x0100ffff, 0x0000ffff, 0xff00ffff,
|
||||
0x0001ffff, 0x0101ffff, 0x0001ffff, 0xff01ffff,
|
||||
0x0000ffff, 0x0100ffff, 0x0000ffff, 0xff00ffff,
|
||||
0x00ffffff, 0x01ffffff, 0x00ffffff, 0xffffffff,
|
||||
GGML_TABLE_END()
|
||||
|
||||
|
|
|
@ -3807,14 +3807,14 @@ void ggml_vec_dot_i2_i8_s(int n, float * restrict s, size_t bs, const void * res
|
|||
// __m256i accu = _mm256_setzero_si256();
|
||||
|
||||
// for (int i=0; i<n/32; i++) {
|
||||
// const int8_t* w0 = (const int8_t *)(i2_q8 + x[i*8 + 0]);
|
||||
// const int8_t* w1 = (const int8_t *)(i2_q8 + x[i*8 + 1]);
|
||||
// const int8_t* w2 = (const int8_t *)(i2_q8 + x[i*8 + 2]);
|
||||
// const int8_t* w3 = (const int8_t *)(i2_q8 + x[i*8 + 3]);
|
||||
// const int8_t* w4 = (const int8_t *)(i2_q8 + x[i*8 + 4]);
|
||||
// const int8_t* w5 = (const int8_t *)(i2_q8 + x[i*8 + 5]);
|
||||
// const int8_t* w6 = (const int8_t *)(i2_q8 + x[i*8 + 6]);
|
||||
// const int8_t* w7 = (const int8_t *)(i2_q8 + x[i*8 + 7]);
|
||||
// const int8_t* w0 = (const int8_t *)(i2s_i8s + x[i*8 + 0]);
|
||||
// const int8_t* w1 = (const int8_t *)(i2s_i8s + x[i*8 + 1]);
|
||||
// const int8_t* w2 = (const int8_t *)(i2s_i8s + x[i*8 + 2]);
|
||||
// const int8_t* w3 = (const int8_t *)(i2s_i8s + x[i*8 + 3]);
|
||||
// const int8_t* w4 = (const int8_t *)(i2s_i8s + x[i*8 + 4]);
|
||||
// const int8_t* w5 = (const int8_t *)(i2s_i8s + x[i*8 + 5]);
|
||||
// const int8_t* w6 = (const int8_t *)(i2s_i8s + x[i*8 + 6]);
|
||||
// const int8_t* w7 = (const int8_t *)(i2s_i8s + x[i*8 + 7]);
|
||||
|
||||
// __m256i xq8 = _mm256_set_epi8(
|
||||
// w0[0], w0[1], w0[2], w0[3],
|
||||
|
@ -3860,7 +3860,7 @@ void ggml_vec_dot_i2_i8_s(int n, float * restrict s, size_t bs, const void * res
|
|||
int sumi = 0;
|
||||
|
||||
for (int i = 0; i < n / 4; i++) {
|
||||
const int8_t* weight = (const int8_t *)(i2_q8 + x[i]);
|
||||
const int8_t* weight = (const int8_t *)(i2s_i8s + x[i]);
|
||||
sumi += (int)y[i*4+0] * weight[0];
|
||||
sumi += (int)y[i*4+1] * weight[1];
|
||||
sumi += (int)y[i*4+2] * weight[2];
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue